메뉴
Structured visual QA: screenshot → batch issues → fix all → verify. Replaces the 300-cycle screenshot→edit death spiral. Optional bishkek review as exit gate. Use when building/polishing UI with browser testing, or when user asks for N iterations/reviews.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
SOC 직업 분류 기준
Behavioral testing methodology — test what users experience, not how code is structured. Use when writing tests, reviewing test quality, planning test strategy for new features, or when existing tests are brittle/verbose/coupled to implementation details. Triggers: writing tests, TDD, test review, "tests keep breaking", "too many mocks", "tests are verbose", test coverage planning, behavior-driven development.
Summon 8–12 fictional but statistically plausible people from Kyrgyzstan to judge your web app. They take screenshots, click around, get confused, get delighted, and write honest reviews in Russian. Output: a single .md report with embedded screenshots. Use when you need brutal UX feedback from people who ride marshrutkas.
Artemy Lebedev-style бизнес-линч design review. Opens the site in a real browser, screenshots everything, then tears it apart (or praises it) in his signature voice — brutal, specific, visual. Use when you want a no-bullshit design review, 'линч', 'lebedev review', 'artemy review', or 'roast my site'.
Autonomous experiment loop — modify code, measure, keep/discard, repeat forever. Based on Karpathy's autoresearch pattern. Use when there's working code + a measurable metric to optimize. Agent works while you sleep.
Iterative deep research agent — recursive breadth×depth tree search that accumulates learnings, generates follow-ups, and produces cited reports. Uses exa + serper + subagents. Use when user says 'deep research', 'investigate', 'deep dive', 'thorough analysis', or needs multi-source synthesis with depth.
Detect and fix design system leaks — default shadcn/bootstrap/MUI styling breaking through your brand. Use when: UI feels 'template-y', 'looks like shadcn', 'too generic', buttons are inconsistent colors, cards look default, or after shipping many features fast. Also: 'audit my design', 'check consistency', 'brand leak', 'fix the defaults'.
| name | visual-qa-loop |
| description | Structured visual QA: screenshot → batch issues → fix all → verify. Replaces the 300-cycle screenshot→edit death spiral. Optional bishkek review as exit gate. Use when building/polishing UI with browser testing, or when user asks for N iterations/reviews. |
Prevents the screenshot→fix→screenshot→fix death spiral (300+ cycles seen in real sessions).
Screenshot every view before touching code. Desktop + mobile. Light + dark. Empty + populated. All major routes.
Look at ALL screenshots. Write one issue list:
### P0 — Broken
- [ ] Login button does nothing (home, mobile)
### P1 — Significant
- [ ] Cards truncated (search, desktop)
### P2 — Polish
- [ ] Spacing inconsistent (search, mobile)
Fix all issues before taking another screenshot. Group by file — one edit per file per round, not per issue.
>5 edits to the same file = stop and refactor it. You're patching a monolith. Split the component first, then fix.
Re-screenshot same views. Check off fixed issues. If regressions appeared, add to list and do one more batch.
Max 3 rounds. If round 4 is needed, the architecture is wrong — refactor, don't patch.
When the list is mostly ✅, optionally run bishkek-report as final QA. One review → one batch fix → done. Not review→fix→review→fix.